On Robustness of Three Linear Time Series Models to an Outlier

Abstract

This article compared the robustness of decomposition, Holt-Winters and SARIMA models to an outlier. Data are average monthly minimum and average monthly maximum temperature data for Bida, collected from the National Cereals Research Institute, Badeggi, Bida, Nigeria. Original series and series with an outlier were analyzed via the three models in contention. Mean absolute percentage error was the performance criterion used.   Out-of-sample forecast accuracy was used as a proxy for robustness. Results revealed Holt-Winters and decomposition as the most robust for minimum and maximum temperature data respectively. SARIMA proved to be the least robust for both series. For the minimum temperature series, decomposition followed Holt-Winters while for the maximum temperature series, Holt-Winters trailed behind decomposition. More series may be required to know the robust of more decomposition and Holt-Winters. It was suggested that more series be included in future studies.